Publication: t-HYPER NEAT: RETARDOS DE TIEMPO EN UNA RED HYPER NEAT PARA APRENDIZAJE DE CAMINATAS EN ROBOTS CON EXTREMIDADES MÓVILES
Date
2017
Authors
SILVA MUÑOZ, OSCAR ANDRE
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
La generación de caminatas en robots con extremidades móviles es una importante tarea parapermitir el correcto desplazamiento de robots en distintos escenarios, ya que el diseño manual deestas involucra un gran desafío debido al elevado uso de tiempo y procesamiento. En este proyectode memoria se ha investigado sobre un popular método de neuroevolución utilizado para esta tarea,llamado HyperNEAT, Hipercubo basado en el Aumento de Topolog´ias. HyperNEAT es un generadorde codificaciones que evoluciona redes neuronales artificiales haciendo uso de los principios del algoritmode Neuroevoluci´on basado en el Aumento de Topolog´ias (NEAT). Es una novedosa t´ecnicapara la evoluci´on de redes neuronales de gran escala usando las regularidades geom´etricas del problemadescritas por la red. En base a este m´etodo es que esta memoria propone la implementaci´ondel m´etodo llamado t-HyperNEAT, incorporando conceptos temporales en una red neuronal HyperNEAT,incluyendo retardos de tiempo adicionales a los pesos en las conexiones entre neuronas,ypermitiendo as´i la generaci´on de caminatas con comportamientos y caracter´isticas m´as cercanas acaminatas vistas en la naturaleza. Posterior a la implementaci´on de ambos m´etodos, estos son puestosa prueba en la generaci´on de caminatas de dos robots con distinto n´umero de grados de libertad.El an´alisis comparativo de los resultados revela que, con respecto a las variables cuantitativas delexperimento, no existe una diferencia relevante en el desempe˜no obtenido entre HyperNEAT y -HyperNEAT, pero no as´i desde el punto de vista cualitativo, obteni´endose notorias diferencias enla coordinaci´on de los movimiento de las extremidades de cada robot, produciendo caminatas m´asnaturales y complejas.
Gait generation for legged robots is an important task to allow an appropiate displacement indifferent scenarios, since hand-tuning design involves a big challenge due to the high computationalefforts translated into processing power and time. This project study a popular neuroevolutionmethod called HyperNEAT, which stands for Hypercube-based Neuroevolution of Augmenting Topologies.HyperNEAT is a generative encoding mechanism which evolves artificial neural networksusing the principles of the algorithm of Neuroevolution of Augmenting Topologies (NEAT). It isa new technique for evolve large scale artificial neural networks using geometric regularities ofthe problem present in the network. Based on this method, this report proposes the implementationof the method called t-HyperNEAT, incorporating temporal concepts in a HyperNEAT neuralnetwork, introducing additional time delays in neural connections between neurons, allowing gaitgeneration with behaviours and characteristics more similar to gaits found in nature. Subsequent tothe implementation of both methods, these are tested in gaits generation of two robots with differentnumber of degrees of freedom. The comparative analysis of the results reveals that, regardingquantitative variables of the experiment, there is no significant difference in performance obtainedbetween HyperNEAT and t-HyperNEAT, but not from the qualitative point of view, obtaining notabledifferences in coordinating the movement of the limbs of each robot, producing more naturaland complex gaits.
Gait generation for legged robots is an important task to allow an appropiate displacement indifferent scenarios, since hand-tuning design involves a big challenge due to the high computationalefforts translated into processing power and time. This project study a popular neuroevolutionmethod called HyperNEAT, which stands for Hypercube-based Neuroevolution of Augmenting Topologies.HyperNEAT is a generative encoding mechanism which evolves artificial neural networksusing the principles of the algorithm of Neuroevolution of Augmenting Topologies (NEAT). It isa new technique for evolve large scale artificial neural networks using geometric regularities ofthe problem present in the network. Based on this method, this report proposes the implementationof the method called t-HyperNEAT, incorporating temporal concepts in a HyperNEAT neuralnetwork, introducing additional time delays in neural connections between neurons, allowing gaitgeneration with behaviours and characteristics more similar to gaits found in nature. Subsequent tothe implementation of both methods, these are tested in gaits generation of two robots with differentnumber of degrees of freedom. The comparative analysis of the results reveals that, regardingquantitative variables of the experiment, there is no significant difference in performance obtainedbetween HyperNEAT and t-HyperNEAT, but not from the qualitative point of view, obtaining notabledifferences in coordinating the movement of the limbs of each robot, producing more naturaland complex gaits.
Description
Catalogado desde la version PDF de la tesis.
Keywords
APRENDIZAJE DE CAMINATAS , HYPERNEAT , NEUROEVOLUCION , RED NEURONAL ARTIFICIAL